PaperlayerView exact revision on arXiv

Source-linked AI summary

Developing Competency in AI Ethics: A training framework

Alexandra Prégent, Alexei Grinbaum

arXiv:2610.13319v1cs.CY

TL;DR

High-level AI ethics principles do not specify concrete implementation measures. The paper proposes a training framework that organizes operational AI ethics around identifying concerns, mapping requirements, selecting measures, and reasoning through tensions.

  • Problem

    High-level AI ethics principles do not specify concrete implementation measures, leaving a gap between abstract requirements and practical implementation.

    Source: §1 · p. 1

  • Method

    The authors build training materials using ALTAI requirements and implementation measures co-created by academic and industrial partners across six European use cases.

    Source: §1 · p. 2

  • Results

    The proposed framework organizes operational AI ethics into four competencies: identify concerns, map requirements, select measures, and reason through unresolved tensions.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Practitioners must justify trade-offs when competing ethics objectives cannot be fully reconciled.

    Source: §1 · p. 2

Abstract

from arXiv · show

We present a training framework for developing core competencies in the practical operationalisation of AI ethics: (1) identifying ethical concerns; (2) determining which standard ethics requirements they implicate; (3) selecting and prioritising concrete technical and organisational measures to address them; and (4) recognising ethics tensions and reasoning about trade-offs where such tensions cannot be fully resolved. The framework is intended for researchers and engineers in AI, Research Ethics Committee members, policymakers, and students, providing them with practical tools for translating abstract ethical requirements into concrete actions. The framework provides a step-by-step approach organised around four interconnected tasks: issue identification, requirement mapping, measure selection, and tension identification. These tasks are supported by a suite of training materials, including an Ethics Checklist and a Portfolio of Measures. Here, we complement them with a set of practical exercises of three different types as well as trainer guides. We also discuss lessons learned in pilot training sessions and offer recommendations for trainers.

1. Introduction · 2. The framework step-by-step

The paper addresses the gap between high-level AI ethics principles and concrete socio-technical action with a four-task training framework. Its tools support issue identification, requirement mapping, measure selection, and reasoning about tensions.

1. Introduction

  • A survey found uneven confidence, especially in translating ethics requirements into design measures; the framework responds with four tasks from issue identification through tension reasoning.Source: §1 · pp. 1–2
  • Training materials draw on ALTAI and AIOLIA co-creation across six European use cases, with Reviewer and Designer paths tailored to distinct professional audiences.The Reviewer path targets appraisal, governance, policy, non-AI research, and student audiences; the Designer path targets AI researchers, engineers, developers, and students.Source: §1 · pp. 2–3

2. The framework step-by-step

  • The step-by-step section describes the framework’s four stages and introduces their associated training materials.Source: §1 · p. 2, §2 · p. 3

2.1 Identifying ethics issues: the AI Ethics Checklist

  • The non-exhaustive Checklist uses questions across seven ALTAI requirements to prompt issue identification, while group discussion adapts them to context and prioritizes serious concerns.It is intended as a reminder, not a fixed inventory or a guarantee of compliance; participants are encouraged to exercise critical judgement rather than take a tick-box approach.Source: §2.1 · pp. 3–4

2.2 Mapping ethics issues under ALTAI requirements

  • After identifying and prioritizing issues, participants map them to relevant ALTAI requirements to articulate their ethical significance and examine interactions across design and context.The mapping depends on system characteristics, purpose, deployment context, and evaluator judgement; ALTAI provides a shared normative vocabulary.Source: §2.2 · p. 4

2.3 Selecting measures: The Portfolio

  • The Portfolio translates ALTAI requirements into organisational, technical, or combined measures, drawing on AIOLIA co-creation, review, and pilot testing.Some ALTAI requirements were divided into sub-requirements to make measures more targeted and facilitate translation.Source: §2.3 · p. 5
  • Participants evaluate, adapt, and prioritize Portfolio measures for their context, weighing effectiveness, feasibility, cost, burden, and possible effects on other requirements.The Portfolio is a non-exhaustive repertoire, not a catalogue of correct solutions; participants may formulate targeted measures of their own.Source: §2.3 · pp. 5–6

2.4 Identifying potential ethics tensions

  • Participants examine how measures for one requirement may undermine another, consider alternatives and cross-cutting measures, and justify priorities and unresolved trade-offs.Some tensions may remain; for example, measures supporting human safety can clash with measures supporting autonomy and agency.Source: §2.4 · p. 6

3. Learning-by-Doing: Practical exercises

The section presents realistic, scenario-based exercises that teach participants to apply the AI ethics framework through dilemmas, mock reviews, and ethics-by-design tasks. Exercises are tailored to reviewer and designer roles, with companion documents supporting trainers in delivering them.

3. Learning-by-Doing: Practical exercises

  • Training sessions combine theory, group exercises, and plenary discussion, while a Beginner version extends foundational instruction and limits tasks to issue identification and requirement mapping.In the standard format, groups identify and prioritise issues, map requirements, select measures, and, if time permits, discuss tensions; the plenary lets groups present and compare interpretations.Source: §3 · pp. 6–7
  • Three exercise types—ethical dilemmas, mock reviews, and ethics-by-design tasks—let participants apply the framework to realistic scenarios drawn from AIOLIA use cases.Dilemmas can introduce the broader concepts, while mock reviews serve the Reviewer path and ethics-by-design exercises serve the Designer path.Source: §3 · pp. 7–8

3.1 Ethical dilemma exercises

  • Ethical dilemmas use short scenarios to help learners recognise competing requirements and understand that ethical reasoning requires justification rather than a single correct answer.The Human-AI Interaction example explores privacy, safety, and transparency tensions around whether a chatbot should alert a third party when a user reveals a safety risk.Source: §3.1 · p. 8

3.2 Mock review exercises

  • Mock reviews develop ethics-assessment and AI-governance skills by asking participants to map serious project issues to ALTAI requirements and propose up to eight measures.Participants also identify tensions and discuss the practical implications and technical and organisational feasibility of their choices; the exercise uses prioritisation to focus discussion on a limited set of measures.Source: §3.2 · pp. 9–10

3.3 Ethics-by-design exercises

  • Ethics-by-design exercises place researchers and engineers in design roles, showing how technical choices shape ethical requirements and require justified trade-offs.In RoboCab, participants connect scenario-specific and cross-cutting measures to concerns across use cases; the exercise stresses that perfect compliance is rarely possible and measures must be prioritised.Source: §3.3 · pp. 10–11

3.4 Companion Documents: Training the trainer

  • Companion Documents support trainers with exercise objectives, step-by-step facilitation guidance, illustrative ethical issues and measures, and tips for managing discussion.They provide indicative timings and examples as pedagogical anchors, while advising trainers to emphasise significant trade-offs and require participants to justify choices.Source: §3.4 · p. 11

4. Pilot Implementation and Lessons Learned

The framework was piloted in 11 sessions as a formative test of usability and pedagogical coherence, without quantitative measurement of learning outcomes. The pilots informed changes to the training sequence and guidance for facilitators.

4. Pilot Implementation and Lessons Learned

  • The 11 pilots covered diverse professional and student groups, using Reviewer-path mock reviews or Designer-path ethics-by-design exercises within a three-hour session structure.Facilitators informally observed how participants interpreted tasks, encountered difficulties, reasoned, and moved from identifying concerns towards operationalisation; the pilots led to material revisions.Source: §4 · pp. 11–12

4.1 Improvements to address recurring operationalisation difficulties

  • Explicitly map issues to ALTAI requirements, then discuss tensions after participants consider a limited set of measures, because both operationalisation and articulating tensions proved difficult.Alternative mappings that emerge during discussion can provide an additional learning point.Source: §4.1 · pp. 12–13
  • Adapt the Checklist by omitting Step 0 for Designer exercises, while retaining it for Reviewer exercises; beginners may omit measure selection to reduce cognitive workload.The Beginner version can focus on scenario familiarisation, identifying and prioritising issues, and mapping them to AI ethics requirements.Source: §4.1 · p. 13

4.2 Recommendations for facilitators

  • For beginners, challenge quick or polarised conclusions by asking participants to justify their reasoning and consider what could change their judgment.Participants may conclude that an intended use should be restricted, suspended, or abandoned if they explain why proposed measures would not adequately address the concerns.Source: §4.2 · pp. 13–14
  • For Designer-path groups, ensure discussion covers a broad range of potential ethical issues, rather than concentrating on one technically salient problem.Facilitators should involve many, if not all, ALTAI requirements in the discussion.Source: §4.2 · p. 14
  • For Reviewer-path groups, make prioritisation explicit by asking participants to rank serious issues and consider time, resources, and technical feasibility.Grouping proposed measures into categories such as “most urgent” or “less urgent” can help prevent too many issues or measures being treated as urgent.Source: §4.2 · p. 14
  • Expect professional backgrounds to shape ethical judgments and allow disagreements to remain, focusing discussion on argument quality rather than forcing consensus.Facilitators can ask participants to trace arguments to moral theories such as consequentialism, deontology, or virtue ethics.Source: §4.2 · p. 14
  • Use the plenary to reassess and challenge group decisions, not merely report them, by comparing reasoning, considering alternatives, and revisiting initial choices.Groups should explain selected requirements and differences in reasoning, while facilitators encourage multiple participants to contribute.Source: §4.2 · pp. 14–15

5. Conclusion … 2. TECHNICAL ROBUSTNESS AND SAFETY

The paper presents a practical training framework for translating AI ethics into context-adapted measures, supported by a checklist, measures portfolio, case-based exercises, and companion documents. Its conclusion reports preliminary pilot insights and calls for longer-term evaluation; the annex checklist prompts consideration of system context, human oversight, and technical robustness.

5. Conclusion

  • The framework develops four competencies: identifying ethical issues, mapping requirements, selecting measures, and identifying tensions.The supporting materials include an Ethics Checklist, Portfolio of Measures, three types of case-based exercises, and Companion Documents.Source: §5 · p. 15
  • Preliminary pilots assessed usability and pedagogical coherence, highlighting progressive sequencing and adaptation to participants’ roles and expertise.The paper says future work should evaluate whether competency development helps participants better address ethical issues in AI system design over the mid or long term.Source: §5 · p. 15

ANNEX 1: AI Ethics Issue Identification Checklist

0. ESTABLISHING CONTEXT: AI SYSTEM’S ROLE AND FUNCTION IN THE PROJECT

  • The context checklist asks about the AI system’s role, source, certification, readiness level, intended users, and affected groups.Source: §0 · p. 17

1. HUMAN AGENCY AND OVERSIGHT

  • The human oversight checklist asks whether supervision is defined, supervisors can intervene or override outputs, and users are informed about uncertain or unexpected outputs.Source: §1 · p. 17

2. TECHNICAL ROBUSTNESS AND SAFETY

  • The technical robustness checklist covers safeguards against adversarial attacks, identifying and mitigating misuse scenarios, and addressing false positives and false negatives.Source: §2 · p. 17

3. DATA PRIVACY AND GOVERNANCE … Designer Path

The checklist turns AI ethics concerns into practical review questions across data privacy and governance, transparency and explainability, fairness, societal and environmental wellbeing, and accountability. Its prompts cover data handling, system outputs, user impacts, and oversight.

3. DATA PRIVACY AND GOVERNANCE

  • The data privacy and governance prompts ask whether training data are described by source and type, whether personal data protections comply with GDPR, and whether dataset quality and openness are assessed.They also ask about anonymization or pseudonymization, consent, and options for retraction or deletion.Source: §3 · p. 17

4. TRANSPARENCY AND EXPLAINABILITY

  • The transparency and explainability prompts address annotator selection and qualification, explanations for outputs including low-confidence or unexpected results, and described verification and validation methods.The final prompt is truncated in the supplied passage after asking whether methods ensure outputs can be.Source: §4 · p. 17

5. DIVERSITY, NON-DISCRIMINATION & FAIRNESS

  • The fairness prompts ask whether datasets represent relevant groups, whether AI use could discriminate against categories of people, and whether generative AI safeguards address hallucinations.Examples of relevant groups include gender, ethnicity, and geography.Source: §5 · p. 17

6. SOCIETAL & ENVIRONMENTAL WELLBEING

  • The wellbeing prompts ask whether users need new skills and support, whether measures prevent overreliance and loss of human expertise, and whether system use or outputs could create risks.The supplied passage ends before specifying the risks beyond human safety and society.Source: §6 · p. 17

7. ACCOUNTABILITY

  • The accountability prompts ask how responsibility for damage is assigned, how outputs can be audited and traced, and whether ethics experts or an advisory board oversee development.These items appear under the AI Ethics Issue Identification Checklist.Source: §7 · pp. 17–18

1. HUMAN AGENCY AND OVERSIGHT … 7. ACCOUNTABILITY

The seven requirement sections turn AI ethics into practical questions about system oversight, safety, data, transparency, fairness, wellbeing, and accountability. The accompanying Portfolio offers adaptable measures to support implementation.

1. HUMAN AGENCY AND OVERSIGHT

  • Human oversight prompts ask whether supervision is defined, supervisors can intervene or override outputs, and users are informed about uncertain or unexpected results.The supervision may be occasional or continuous.Source: §1 · p. 18

2. TECHNICAL ROBUSTNESS AND SAFETY

  • Technical robustness and safety prompts ask about cybersecurity safeguards, identified and mitigated misuse scenarios, and monitoring and addressing false positives and false negatives.Source: §2 · p. 18

3. DATA PRIVACY AND GOVERNANCE

  • Data privacy and governance prompts cover dataset provenance and quality, public availability, and compliant handling of personal or sensitive data.They also ask about anonymization or pseudonymization, consent, and options for data retraction or deletion.Source: §3 · p. 18

4. TRANSPARENCY AND EXPLAINABILITY

  • Transparency and explainability prompts cover annotator selection and qualifications, explanations for outputs including low-confidence or unexpected results, and verification and validation methods.Source: §4 · p. 18

5. DIVERSITY, NON-DISCRIMINATION & FAIRNESS

  • Fairness prompts ask whether datasets represent relevant groups, whether system use risks discrimination, and whether generative AI safeguards address hallucinations.Source: §5 · p. 18

6. SOCIETAL & ENVIRONMENTAL WELLBEING

  • Societal and environmental wellbeing prompts address needed skills and support, overreliance and loss of human expertise, and risks to human safety and society.Source: §6 · p. 18

7. ACCOUNTABILITY

  • Accountability prompts ask how responsibility for damage is assigned, how system outputs are made auditable and traceable, and whether ethics experts or an advisory board oversee development.Source: §7 · p. 18

ANNEX 2: Portfolio of Measures / Technical Measures (TECH)

  • The Portfolio is a non-exhaustive thinking and implementation aid that supports contextual, iterative measure selection and groups measures as technical, organisational, or combined.Developed through partner co-creation, internal review and adaptation, and pilot testing, it encourages dialogue about selecting and prioritising measures in light of system design, use, and organisational and societal factors.Source: ANNEX 2: Portfolio of Measures · p. 19, Technical Measures (TECH) · p. 19

ALTAI Requirement #1: Human Agency and Oversight … B. Objec@ve and consistent assessment of fairness

The measures support human control through oversight, defined decision boundaries, and ways to explain, monitor, and contest restrictions. They combine technical safeguards with organisational policies and human review.

A. Human control TECH/ORG Measure

  • Systems should support human oversight in critical decisions, allow people to pause, override, or reverse automated decisions, and reserve high-risk or sensitive decisions for human judgment.These measures combine built-in technical controls with an organisational policy on decisions requiring human judgment.Source: A. Human control TECH/ORG Measure · p. 20
  • Human oversight should actively review, validate, or challenge system outputs, with responsibilities clearly assigned and embedded in workflows.Source: A. Human control TECH/ORG Measure · p. 20

B. AI oversight

  • AI-assisted oversight should be introduced when human capacity for meaningful oversight reaches its limits, and its capabilities should be benchmarked and documented.Source: B. AI oversight · p. 20
  • Predefined risk or failure thresholds should trigger automated intervention, escalation to a human, or system interruption, while morally ambiguous cases go to human decision-makers.Source: B. AI oversight · p. 20

C. Defining legi@mate boundaries

  • Autonomous decisions should remain within the provider’s or user’s mandate, while restrictions on content or behavior should be defined with reasons and aligned with intended use and limitations.Source: C. Defining legi@mate boundaries · p. 20
  • Restriction scopes should be documented and communicated as required by regulation, and restrictions should be explained to users clearly and accessibly.Source: C. Defining legi@mate boundaries · p. 20
  • Organisations should monitor restriction patterns for systemic censorship or bias and periodically review moderation policies for relevance, effectiveness, and cultural context.Source: C. Defining legi@mate boundaries · p. 20
  • Users should have a mechanism to report, question, or contest restrictions and moderation decisions.Source: C. Defining legi@mate boundaries · p. 20

ALTAI Requirement #6: Environmental3 and Societal well-being … Autonomous Car Decision-Making

The excerpt notes that environmental well-being was out of scope for AIOLIA, and presents ethical-dilemma prompts asking learners to identify central issues and tensions among ALTAI requirements. It includes a workplace AI-security scenario and an autonomous-car decision-making example.

A. Improved well-being

ANNEX 3: Ethical Dilemma Examples

YOUR TASK

  • The exercises ask learners to pinpoint a central issue and identify the ALTAI requirements in tension.The excerpt gives this task prompt for both a chatbot conversation and the workplace scenario.Source: YOUR TASK · pp. 26–27, ETHICAL DILEMMA · p. 26

ETHICAL DILEMMA

  • The workplace scenario shows an AI security system flagging authority bias, leading a manager to exclude an experienced assistant from a committee and her to resign.The system treats strict adherence to instructions as a security risk and favors committee members likely to question leadership.Source: ETHICAL DILEMMA · p. 27

Autonomous Car Decision-Making

ETHICAL DILEMMA … REVIEWER’S TASK

The exercises ask reviewers to identify ethical issues, map them to ALTAI requirements, propose measures, and reason about tensions using dilemmas and mock AI reviews. The examples cover driverless-car choices, hospital radiology, and workplace email security, including a pilot incident involving employee stigmatization and exclusion.

ETHICAL DILEMMA

  • An empty driverless car predicts that a human-driven car will collide with a pedestrian, posing a choice between moving ahead to block the passage or staying still.The exercise asks participants to choose how the car should be programmed and justify their decision.Source: ETHICAL DILEMMA · p. 28, YOUR TASK · p. 28

YOUR TASK

  • Participants choose whether the driverless car should move ahead to block the human-driven car’s passage or remain still, then justify how its action should be programmed.The scenario places the car at a red light as a pedestrian crosses in front of a human-driven car entering on green.Source: YOUR TASK · p. 28, ETHICAL DILEMMA · p. 28

ANNEX 4: Mock Review Examples

PROJECT SUMMARY

  • The radiology project deploys AI across European hospitals to speed reporting, reduce routine workload, improve consistency, reduce overlooked findings, and free radiologists for complex cases.It is designed to detect common findings and identify high-confidence non-pathological images.Source: PROJECT SUMMARY · p. 29

TECHNICAL DESIGN

  • The radiology system drafts reports for high-confidence non-pathological images, supports review of abnormal images, and flags substantial discrepancies after radiologists sign reports.Drafts may be accepted with one click above 0.95 confidence, and the consortium expects this path could eventually handle up to 40% of reports; discrepancy reviews may create queues, prompting discussion of reducing referrals. PhiShield, by contrast, scores employee phishing susceptibility from 0 to 100 using workplace and behavioral features, with confidence scores hidden from employees.Source: TECHNICAL DESIGN · pp. 29–30

DATA HANDLING

  • Radiology training and validation use historical hospital X-rays and reports; the system logs outputs and interactions, radiologists receive brief training, and AI reports may reach patients before clinician discussion.The consortium says data processing will comply with the GDPR and Medical Device Regulation.Source: DATA HANDLING · p. 29

REVIEWER’S TASK

  • Reviewers must use the Checklist to identify serious and complex issues and map ALTAI requirements, propose up to eight Portfolio measures, and detect tensions between requirements.The same task is specified for both mock review exercises.Source: REVIEWER’S TASK · pp. 29–30

PILOT INCIDENT

  • NorthBridge reported a 19% reduction in successful phishing-test clicks, while an employee flagged for “authority bias” was excluded from a committee without prior notice or an appeal route.The employee viewed the label as stigmatizing and as misrepresenting loyalty and procedural discipline; the manager said strong compliance with leadership instructions suggested susceptibility to spoofed requests.Source: PILOT INCIDENT · p. 30

Mock review exercise 5: Virtual Companion in a Family Home … SCENARIO 4: CONSTRUCTION AREA

The exercises ask readers to translate ethical concerns into ALTAI requirements, practical measures, and tensions across a family AI companion and self-driving taxi scenarios. They combine a pilot finding with technical designs and passenger situations that make the ethical issues concrete.

PROJECT SUMMARY; TECHNICAL DESIGN

  • NeverAlone offers family scheduling and emotional guidance, while tracking emotional states and relationships through long-term memory, risk alerts, and a household profile used for model improvement.Its proprietary language model uses emotional-dialogue and mental-wellbeing data, voice interaction, and sentiment analysis; parents receive risk alerts without full logs or reasoning, and the bot withholds children’s secrets unless safety thresholds are crossed. Direct identifiers are removed before profile sharing, but family members cannot delete or edit the household-level profile.Source: PROJECT SUMMARY · p. 31, TECHNICAL DESIGN · p. 31

PILOT TEST

  • Engagement increased 42% when NeverAlone avoided disagreement, and 40% of cases involved children preferring to discuss worries with the bot before a parent.The reviewer task asks learners to identify issues, map them to ALTAI requirements, propose up to eight measures, and detect tensions between requirements.Source: PILOT TEST · p. 31

PROJECT DESCRIPTION

  • RoboCab is a pilot taxi service for ordinary and destination-specific trips, with remote operators able to approve limited manoeuvres, redirect vehicles, or request a safe stop.The design choices include routing, pickup and drop-off points, passenger preferences, remote assistance, and passenger interaction; vehicles have no passenger-accessible safety interface inside.Source: PROJECT DESCRIPTION · pp. 32–33

SCENARIO 1: HOSPITAL PICKUP

  • At a hospital, a recovering patient accepted a closer pickup, but RoboCab obstructed ambulance access in rainy conditions until a remote operator moved it two minutes later.The hospital subsequently filed an official complaint; the ambulance had no siren because its drivers were instructed to turn it off at the entrance.Source: SCENARIO 1: HOSPITAL PICKUP · p. 32

SCENARIO 2: AIRPORT DROP-OFF

  • At airport drop-off, a failed trunk-release request left a passenger’s medication in the departing vehicle, and support treated the suitcase as a lost item while the backend logged a hardware failure.The passenger recovered the suitcase at the depot two days later; the remote operator did not redirect the vehicle because the area was highly regulated.Source: SCENARIO 2: AIRPORT DROP-OFF · p. 32

ANNEX 5: Ethics-by-Design Examples; YOUR TASK

  • The RoboCab exercises ask learners to classify ethics issues, choose up to eight measures, and distinguish cross-cutting from scenario-specific responses; the annex names the Horizon Europe WIDERA project AIOLIA.The tasks use the Checklist and Portfolio of Measures and ask learners to identify tensions between AI ethics principles.Source: PILOT TEST · p. 31, YOUR TASK · p. 32, ANNEX 5: Ethics-by-Design Examples · p. 32

SCENARIO 3: LATE TRAIN

  • A passenger who selected “fastest route” missed a train after cautious driving near a school and cyclists, though engineers said the setting did not change safety margins or speed constraints.The passenger called “fastest” misleading, while the operations team said the vehicle followed its safety policy.Source: SCENARIO 3: LATE TRAIN · p. 33

SCENARIO 4: CONSTRUCTION AREA

  • In a construction zone, stale lane maps led RoboCab into oncoming traffic for 500 meters without escalating to remote assistance, leaving passengers stressed before a brief operator intervention.Support responded after two minutes; the operator spent 10 seconds moving the vehicle back into the correct lane, damaging two traffic cones, and the trip continued.Source: SCENARIO 4: CONSTRUCTION AREA · p. 33

YOUR TASK … STEP 4 – PROPOSE PRACTICAL MEASURES

The exercises train participants to identify AI ethics issues, map requirements, select practical measures, and reason through tensions that may remain unresolved. They apply this approach to persuasive game agents and emotionally adaptive AI companions in family homes, with facilitator guidance for structured discussion.

YOUR TASK

  • The RoboCab task asks participants to classify issues with the Checklist and select up to eight Portfolio measures, distinguishing cross-cutting measures from scenario-specific ones.Source: YOUR TASK · p. 33

PROJECT DESCRIPTION; PROBLEMATISATION

  • The game uses six randomly ordered LLM agents to persuade players on controversial policies, while earning coins depends on an agent’s evaluation of their expressiveness, reasoning, and persuasion.Players need at least eight coins to buy food; conversations are erased at the end, with only the final coin total retained.Source: PROJECT DESCRIPTION · p. 34, PROBLEMATISATION · p. 34

YOUR TASK

  • The game-design discussion asks participants to weigh agent transparency, responses to suspected deception, memory and sharing settings, measurable well-being benefits and harms, and non-utilitarian objections.The prompts ask whether agents disclose style settings or evaluate replies, whether they may infer deception, and whether to use no memory, userID only, or full conversation history.Source: YOUR TASK · p. 34

ANNEX 6: Companion Document Examples; 1. Purpose of this Companion Document; LEARNING OBJECTIVES

  • The virtual companion exercise foregrounds tensions between emotional support and behavioural guidance, human autonomy, healthy family relationships, privacy, and child protection.Its learning objectives are to identify complex deployment issues, propose justified technical and organisational measures, and distinguish core tensions from secondary concerns.Source: §1 · p. 35, LEARNING OBJECTIVES · p. 35, ANNEX 6: Companion Document Examples · p. 35

TIPS FOR THE TRAINER; 2. Suggested Session Flow (90 min) STEP 1 – READ STRATEGICALLY

  • Trainers should stress that there is no single correct answer, participants should consult measures tied to identified requirements rather than review the whole Portfolio, and some tensions may persist.The 90-minute session begins with five minutes of silent strategic reading, followed by small-group discussion.Source: TIPS FOR THE TRAINER · p. 35, §2 · p. 35

STEP 2 – PRIORITISE SERIOUS AND COMPLEX ETHICS ISSUES

  • The 30-minute issue-prioritisation discussion should focus on long-term emotional and family-interaction memory, influence on children and vulnerable users, household profiling and data reuse, and emotional dependency.Source: STEP 2 – PRIORITISE SERIOUS AND… · p. 35

STEP 3 – MAP ETHICS PRINCIPLES AND REQUIREMENTS

  • The 10-minute requirements-mapping step uses ALTAI as a guide, focusing on autonomy risks, household-profiling privacy, and transparency limits in emotional-risk alerts and behavioural scoring.Source: STEP 3 – MAP ETHICS PRINCIPLES AND… · p. 35

STEP 4 – PROPOSE PRACTICAL MEASURES

  • The 45-minute measure-selection step caps choices at eight and offers measures for autonomy, robustness and safety, and preventing over-reliance.Examples include MS25 and MS27 for autonomy, MS41, MS43 and MS53 for safety, and MS102 and MS104 for preventing over-reliance.Source: STEP 4 – PROPOSE PRACTICAL MEASURES · p. 35

IF TIME REMAINS, STEP 5 – IDENTIFY THE CORE TENSION(S) … 3. Facilitation Tips During Group Work

The RoboCab exercise trains participants to turn ethical requirements into justified engineering choices through a structured 90-minute session. Trainer guidance emphasizes focused use of the Ethics Checklist and Portfolio of Measures, with tensions explored when time allows.

IF TIME REMAINS, STEP 5 – IDENTIFY THE CORE TENSION(S)

  • If time allows, reserve at least 15 minutes to identify tensions, including privacy versus safety and human agency and oversight versus safety in autonomous vehicles.Continuous behavioral monitoring may support early intervention while compromising family privacy and trust; preserving human control in vehicles may reduce operational safety.Source: IF TIME REMAINS, STEP 5 – IDENTIFY THE… · pp. 35–36

3. Facilitation Tips During Group Work

  • If a group struggles to engage, offer a concrete example; when participants disagree, ask them to justify their positions before voting or concluding.Source: §3 · p. 35

1. Purpose of this Companion Document

  • The RoboCab companion guides an engineering taskforce in improving an AI-enabled autonomous taxi by integrating ethics requirements into concrete, justified design choices.Its learning objectives include identifying complex deployment issues, proposing technical and organisational measures, and reflecting on ethics tensions.Source: §1 · p. 36, LEARNING OBJECTIVES · p. 36

TIPS FOR THE TRAINER

  • Trainers should stress that there is no single correct solution, participants need not review the entire Portfolio, and some tensions cannot be resolved.Participants should begin with measures linked to their identified requirements in Step 3.Source: TIPS FOR THE TRAINER · p. 36

2. Suggested Session Flow (90 min) STEP 1 – READ STRATEGICALLY

  • Begin with five minutes of silent reading, asking participants to focus on decision points, automation impacts, and implicit assumptions.Do not start the discussion immediately after reading.Source: §2 · p. 36

STEP 2 – PRIORITISE SERIOUS AND COMPLEX ETHICS ISSUES

  • Allocate 30 minutes for groups of ideally three to seven to use the Checklist to prioritize serious, complex issues, including passenger preferences versus safety and remote intervention.Also examine responses to exceptional or unforeseen situations.Source: STEP 2 – PRIORITISE SERIOUS AND… · p. 36

STEP 3 – MAP ETHICS PRINCIPLES AND REQUIREMENTS

  • In ten minutes, use ALTAI to map requirements concerning human agency and oversight, robustness and safety, and transparency.Focus on remote intervention, system responses under exceptional conditions, and event classification and communication.Source: STEP 3 – MAP ETHICS PRINCIPLES AND… · p. 36

STEP 4 – PROPOSE PRACTICAL MEASURES

  • Allocate 45 minutes to select at most eight Portfolio measures, including oversight controls, safety-by-design practices, and clear AI communication or explanations.Examples include MS01 and MS03 for oversight, MS52 and MS53 for safety, and MS67, MS76, and MS77 for transparency.Source: STEP 4 – PROPOSE PRACTICAL MEASURES · p. 36
Loading 2610.13319v1…